Hitzen Adiera-desanbiguazioa

Oier López de Lacalle, Eneko AgirreEKAIA 2010 · DOI

Abstract

Human language is inherently ambiguous. A single word can take on multiple meanings depending on its context, a phenomenon humans navigate naturally but which poses a significant challenge for computational systems. Word Sense Disambiguation (WSD) is the process of automatically assigning the correct sense to a word occurrence using computational methods. This task draws on various knowledge sources, ranging from sense-tagged corpora to complex ontologies. However, the creation of these resources is labor-intensive, leading to what is known as the "knowledge acquisition bottleneck." This paper provides a general introduction to WSD, focusing on supervised learning methods applied to both Basque and English. We demonstrate that WSD is a complex task that requires machines to move beyond simple string processing to perceive the underlying semantic structure of text.

Introduction

Word Sense Disambiguation (WSD) is often described as an "AI-complete" problem. To solve it perfectly, a machine would require human-level intelligence and a comprehensive understanding of world knowledge. Published in the Basque scientific journal EKAIA, the research titled "Hitzen Adiera-desanbiguazioa" surveys the methodologies used to bridge the gap between how humans and machines process language. While the study situates its findings within the global Natural Language Processing (NLP) landscape, it specifically highlights the progress and challenges unique to the Basque language.

The Problem of Linguistic Ambiguity

The primary hurdle in NLP is that machines traditionally process "strings" (sequences of characters) rather than "things" (concepts). For example, the word "bank" can refer to a financial institution, a geographical feature (a river bank), or even a physical object (a bench). While context allows humans to resolve these differences effortlessly, machines require explicit models to distinguish between them. Without effective WSD, downstream technologies such as Machine Translation, Information Retrieval, and the development of the Semantic Web are fundamentally limited.

Methodological Framework: Supervised WSD

The paper focuses heavily on Supervised WSD, currently the most successful approach. In this framework, models are trained on corpora where human experts have manually tagged words with their correct meanings. The process involves a sophisticated pipeline of feature engineering and algorithmic classification.

Feature Engineering

To understand context, models extract four primary types of features from the text:

  • Local Features: Examining the immediate neighbors of a word (e.g., bigrams and trigrams).
  • Topical Features: Analyzing the "bag of words" within the broader document to identify the general subject matter.
  • Syntactic Features: Identifying grammatical relationships, such as subject-verb-object structures.
  • Semantic Features: Incorporating domain-specific information to narrow down potential meanings.

Learning Algorithms

Several machine learning approaches are evaluated for their effectiveness in classifying these features:

  • Naïve Bayes: Used as a probabilistic baseline.
  • k-Nearest Neighbors (k-NN): A "lazy learning" method that classifies new word instances by comparing them to the most similar examples in the training data.
  • Support Vector Machines (SVM): A high-performing algorithm that identifies the optimal hyperplane to separate different word senses in a multi-dimensional space.
  • Ensemble Methods: Techniques like AdaBoost are used to combine multiple weak classifiers into a single, robust system.

Data Resources and Benchmarks

The effectiveness of WSD systems is inextricably linked to the quality of the available data. The paper identifies several critical benchmarks:

  • English Resources: The SemCor corpus and the Wall Street Journal corpus serve as primary training grounds, with WordNet providing the essential "sense inventory" or dictionary of meanings.
  • Basque Resources: The authors highlight the Euskaldunon Egunkaria corpus and the Zientzia eta Teknika Corpusa (ZTC), alongside Euskal WordNet, which are vital for applying these technologies to the Basque language.
  • Evaluation Frameworks: The study references results from the Senseval and SemEval competitions, which standardize testing across "Lexical Sample" (focusing on specific words) and "All-words" (tagging every word in a text) tasks.

Key Insights and Results

The research yields several important conclusions regarding the state of the art in WSD:

  1. The Knowledge Acquisition Bottleneck: The primary limitation of supervised systems is the high cost of manual data labeling. This makes it difficult to scale WSD to every language and specialized domain.
  2. Performance in Basque vs. English: In the Senseval-3 (Basque) task, supervised systems achieved F-measure scores of approximately 70%, with SVM and AdaBoost significantly outperforming the "Most Frequent Sense" baseline of 55.8%. In English (SemEval-2007), accuracy reached 88% for Lexical Sample tasks but dropped to 59% for the more difficult All-words task.
  3. Granularity and Complexity: The difficulty of the task is tied to the "fineness" of the sense distinctions. Distinguishing between vastly different meanings is relatively simple, but separating subtle nuances remains a challenge even for human annotators.
  4. Language Independence: The study confirms that WSD techniques are largely language-independent. Methods developed for English are highly effective for Basque, provided the necessary linguistic resources (like Euskal WordNet) are available.

Conclusion and Future Directions

While supervised methods currently provide the highest accuracy, the field is shifting toward Knowledge-Based WSD. These emerging methods leverage the structural properties of ontologies and graphs to perform disambiguation with less reliance on hand-tagged data. Overcoming the knowledge acquisition bottleneck and improving domain adaptation (e.g., moving from news text to medical text) remain the most critical goals for the next generation of NLP tools. Achieving maturity in WSD will ultimately transform how we access information, paving the way for a truly semantic and interconnected web.

Summary updated

Abstract

Gure hizkuntza anbiguoa da. Hitz batek hainbat interpretazio ditu agertzen den testuinguruaren arabera, eta zein adiera hartzen duen asmatzea ez da lan erraza, nahiz eta guk era naturalean egin. konputazio-metodoak erabiliz hitzen agerpenei adiera egokia ematea hitzaren adiera-desanbiguazioa (HAD) deritzo. HAD automatikoa ezagutzan oinarritzen da, eta ezagutza hori hainbat iturritatik lor dezake: adierez etiketatutako corpus batetik hasita ontologietaraino. Zoritxarrez, baliabide hauen sortze-prozesua garestia da eta ezagutza bereganatzearen arazo bezala ezagutzen da. Eragozpenak gaindituta HAD teknologiak heldutasuna lortzen duen unean informazioa atzitzeko dugun modua erabat aldatuko da, Web Semantikoari ateak zabalduz. Lengoaia Naturalaren Prozesamendurako tresnetan ere lagungarri da HAD automatikoa, Itzulpen Automatikoan adibidez, izan ere, bai polisemia eta bai sinonimia arazoak automatikoki gainditzeko balio baitezake. Artikulu honetan hitzen adiera-desanbiguazioaren sarrera orokor bat egiten dugu eta, bereziki, adierez etiketatutako corpusetan oinarritzen diren metodoei erreparatuko diegu. Erakutsiko dugu hitzen adiera-desanbiguazioa ataza zaila dela, hizkuntzaren konplexutasunari aurre egin eta testu hutsetik egitura semantikoa antzeman behar duelako.
BibTeX
@inproceedings{lacalle2010hitzen,
  author = {Oier López de Lacalle and Eneko Agirre},
  title = {{Hitzen Adiera-desanbiguazioa}},
  booktitle = {EKAIA},
  year = {2010},
  doi = {10.1387/ekaia.1273},
}
Figure